[flang-commits] [libunwind] [clang-tools-extra] [libc] [flang] [lldb] [lld] [compiler-rt] [libcxx] [llvm] [clang] [X86] Use RORX over SHR imm (PR #77964)

[Simon Pilgrim via flang-commits]

================
@@ -4216,6 +4217,97 @@ MachineSDNode *X86DAGToDAGISel::emitPCMPESTR(unsigned ROpc, unsigned MOpc,
   return CNode;
 }
 
+// When the consumer of a right shift (arithmetic or logical) wouldn't notice
+// the difference if the instruction was a rotate right instead (because the
+// bits shifted in are truncated away), the shift can be replaced by the RORX
+// instruction from BMI2. This doesn't set flags and can output to a different
+// register. However, this increases code size in most cases, and doesn't leave
+// the high bits in a useful state. There may be other situations where this
+// transformation is profitable given those conditions, but currently the
+// transformation is only made when it likely avoids spilling flags.
+bool X86DAGToDAGISel::rightShiftUncloberFlags(SDNode *N) {
+  EVT VT = N->getValueType(0);
+
+  // Target has to have BMI2 for RORX
+  if (!Subtarget->hasBMI2())
+    return false;
+
+  // Only handle scalar shifts.
+  if (VT.isVector())
+    return false;
+
+  unsigned OpSize;
+  if (VT == MVT::i64)
+    OpSize = 64;
+  else if (VT == MVT::i32)
+    OpSize = 32;
+  else if (VT == MVT::i16)
+    OpSize = 16;
+  else if (VT == MVT::i8)
+    return false; // i8 shift can't be truncated.
+  else
+    llvm_unreachable("Unexpected shift size");
+
+  unsigned TruncateSize = 0;
+  // This only works when the result is truncated.
+  for (const SDNode *User : N->uses()) {
+    auto name = User->getOperationName(CurDAG);
----------------
RKSimon wrote:

unused variable

https://github.com/llvm/llvm-project/pull/77964